An Automated Search Model (ASM) for the 2D shape has been developed and validated for the 9-,13-, and 17-year-olds, but for the 6-year-old a new ASM had to be created. Here it became obvious that traditional measures of dysplasia using ‘the best fitting circle’ around the femoral head are invalid in this age group given the immaturity of the bones and subsequent non-circular shape of the femoral head. This gave rise to a novel research question on whether the hemispherical acetabulum could be used, rather than the femoral head, to establish the center of rotation of the hip joint (see ‘Novel methodologies’).
The landmark placement and subsequent calculations of morphological measurements (shape) on the DXA images for the 9-13-year-olds have been completed. We also performed the first analyses with regard to the difference in risk factors for hip dysplasia in newborns (DDH) and for hip dysplasia at age 13 years.
One of the first steps in studying 3D shape variations in the hip joint based on medical images is post-processing, particularly the extraction of the anatomical region of interest from the images, known as segmentation. Due to data access challenges in the first phase of the project (~6 months), we used a publicly available dataset instead of starting with MRI scans from the Generation R cohort. This novel pipeline offers several benefits for pediatric medical image processing. Our pipeline demonstrated robust performance in handling growth-related bone changes and addressing image quality variations, including irrelevant objects such as the appearance of a child’s parents' hands on top of their child's hands.
We have recently extended our segmentation pipeline to the semantic segmentation of the pelvis and femurs from MRIs. The strength of our segmentation pipeline lies in its ability to segment MRIs at any age without requiring manually annotated datasets, regardless of the specific age it was trained for. Currently, we are manually segmenting a set of MRIs, including approximately 60 left and right femurs and 30 pelvises, to use as a gold standard for quantitatively evaluating the performance of our segmentation pipeline. Additionally, we are building two 3D Statistical Shape Models (SSMs), one for the pelvis and one for the femur, to study their shape variations across different age groups, specifically 9, 13, and 17 years. Statistical Shape Models (SSMs), one for the pelvis and one for the femur, to study their shape variations across different age groups, specifically 9, 13, and 17 years. In the next phase of the project, we will extend the 3D hip growth models, which represent 3D geometry, to a 5D hip growth model (geometry, density, and time) by leveraging deep-learning-based approaches.
The semiquantitative measurements for joint integrity using the Scoring Hip Osteoarthritis with MRI (SHOMRI) system have been performed in a subset of 576 hips. As it is not feasible to manually score all hips in the dataset, we will build an artificial intelligence (AI)-based model to automatically distinguish between healthy hips and hips with osteoarthritis-related abnormalities (based on SHOMRI scores). We also have developed a model to automatically segment cartilage and bone and extract quantitative measurements for hip dysplasia and joint integrity assessment from pelvic MRI. The algorithm allows for automatic extraction of several morphology measurements including cartilage volume measurements and 3D acetabular coverage of femur.
We have been developing a pipeline to automate the creation of subject-specific finite element models to study the effects of mechanical loads on hip growth. The automation process includes, for instance, the automated determination of muscle attachment points. To achieve this, we have built two 3D Statistical Shape Models (SSMs), one for the pelvis and one for the femur. These models will help identify muscle insertion points for previously unseen individual femurs and pelvises. Another exemplary automation feature is the identification of the region on the femoral head covered by the acetabulum. Additionally, we have implemented a bone remodeling model that relates mechanical loading to changes in the density distributions of relevant bones. We have performed a set of finite element analyses using the publicly available finite element modeling software FEBIO, based on medical images of individuals with and without dysplastic hips. These analyses aim to study the effects of mechanical loading on bone density distributions in both healthy and dysplastic hip cases. Currently, we are post-processing the results of these computational analyses.
